Company · Commitments
Commitments you can audit, not admire.
Principles that live in a PDF protect nobody. Ours are properties of the software, each one checkable in the audit record of any script we have ever marked.
To our customers · to the industry · to the AI itself
01 · To our customers
The record answers, not the vendor.
Four properties of the system. Each one holds in the audit record of any script Sokros has ever marked, and each one can be tested on your own work during a demo.
A learner can always ask why.
Every grade, referral and integrity signal carries verbatim evidence from the learner's own work and written reasoning. 'The model decided' is never the answer. The record is.
Learner data is minimised and bounded.
The pipeline stores decisions and quoted evidence, not profiles. No learner data goes to consumer AI services, retention follows centre instruction, and on-premise or in-region deployment keeps data where the law expects it.
How we handle data →Changes are owned and reversible in the record.
Every behavioural change is a numbered amendment with rationale, replay-tested before release. A release is blocked outright if any calibration script flips between Refer and a clear Pass.
Humans hold the posts that matter.
Assessor-in-the-loop modes (sampling, review-before-release, approval) are first-class features, not workarounds. Centres choose where human judgement sits, and it always sits somewhere.
02 · To the industry
Judged on method, in writing.
How we behave while a centre is still choosing. Three rules for the sales process itself.
Compare methods, not vendors.
When a centre weighs us against another tool, we ask them to compare the marking method: what decides a grade, what evidence sits behind it, what happens on replay. If the other method stands up to those questions, they should buy it.
Compare the fine print →Publish what we can check.
Every claim on this site names what it counts, which is why the claims are narrow. Zero third-party AI in the marking path. Calibration from dozens, not thousands. Narrow claims can be checked; broad ones can only be believed.
Answer in writing.
A due-diligence question gets a written answer that goes on the file, not a call that leaves no trace. If the answer changes later, the correction is written too, with the date it changed.
03 · To the AI itself
What the machine may be.
The engine's own conduct rules: what it may never do, and what it must prove before every release.
Machines never punish.
Integrity outputs (AI-usage indicators, plagiarism signals, reference-credibility flags) are evidence for a qualified human's judgement. No automatic penalty exists anywhere in the system.
The same script gets the same grade.
Fairness starts with consistency. Determinism by architecture means a learner's grade does not depend on the time of day, the queue position, or a sampling seed.
Empirical or absent.
Any behaviour we claim for the engine exists as a replay test, or it is not claimed. Each release re-marks the calibration set and compares the output byte for byte; it is blocked if a single script flips between Refer and a clear Pass. A property of this system is empirical, or it is absent.
04 · The stakes
Because a grade is someone's next step.
Behind every script is a promotion case, a visa condition, a career change. That is why we treat consistency, evidence and human oversight as engineering requirements, the same way aviation treats checklists.
Hold us to it.
Bring any commitment on this page to a 30-minute briefing and ask to see it in the record.
Questions answered in writing · the record open on the call